Your dropout layer zeroes each hidden unit with probability 0.5 during training and leaves the survivors unscaled. At inference you switch dropout off.
Validation loss in eval mode is much worse than with dropout left on, and hidden activations are about twice their training size.